Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (9): 263-270.DOI: 10.3778/j.issn.1002-8331.2010-0383

• Engineering and Applications • Previous Articles     Next Articles

Book Recommendation Based on Heterogeneous Network Representation Learning

ZHANG Jinzhu, JIANG Linqi, WANG Yue, KONG Jie, GAO Yang   

  1. 1.School of Economics and Management, Nanjing University of Science and Technology, Nanjing 210094, China
    2.Nanjing University of Science and Technology, Nanjing 210094, China
  • Online:2022-05-01 Published:2022-05-01

基于异构网络表示学习的相关图书推荐研究

张金柱,蒋霖琪,王玥,孔捷,高扬   

  1. 1.南京理工大学 经济管理学院,南京 210094
    2.南京理工大学,南京 210094

Abstract: This paper plans to expand and integrate various book features and relationships, improves the accuracy and diversity of book recommendation from the semantic-related perspective, and explores the contribution and impact of different book features on recommendation. Firstly, book features are extracted to build the book heterogeneous network, and the multidimensional relationships between features are define. Then, a variety of book features are combined to form the semantic vector representation of the book based on heterogeneous network representation learning, and the related books are recommended by calculating the semantic correction between vectors based on a selected similarity calculation index. Finally, the accuracy of results is evaluated by RMSE and MAE, and the diversity of the results are evaluated by variety, balance and disparity. The Amazon book dataset is extended by multiple features, such as author, keyword and publisher, to construct the book heterogeneous network. The empirical result shows that RMSE and MAE are respectively reduced by 19.52%, 20.51% mostly comparing with the collaborative filtering algorithm. RMSE and MAE also are respectively reduced by 0.17% and 2.9% mostly comparing with the deepwalk. This method can effectively improve the accuracy of recommendation. The indicators of diversity show that this recommend books are more kinds of category and rich in content, this method also improves the diversity. And this paper clarifies the contribution of different features on recommendation, in descending order of contribution: author, keyword, category, user and publisher.

Key words: book recommendation, network representation learning, book heterogeneous network

摘要: 扩展和综合利用多种图书特征及其关联关系,从语义相关的角度提高图书推荐的准确性和多样性,探索不同特征对于图书推荐的贡献程度和影响。抽取多种图书特征构建图书异构网络并设计形成特征间的多维关联关系。引入异构网络表示学习方法,融合多种图书特征,形成图书的语义向量表示,选取向量相似度指标计算并表示图书间的语义相关程度,实现相关图书推荐。利用平均绝对误差、均方根误差等定量指标评估推荐的准确性,利用丰富度、均衡性、差异度等指标分析图书推荐的多样性。扩展亚马逊图书数据集,增加作者、关键词和出版社等特征项构建图书异构网络。实证结果表明,相较于协同过滤,该方法的均方根误差、平均绝对误差最多分别降低了19.52%和20.51%,相较于deepwalk方法,该方法在均方根误差、平均绝对误差最多分别降低了0.17%和2.9%,准确性得到较大提高;多样性评测指标也显示该方法推荐的图书种类更多元、内容更丰富,多样性同样得到了提高;明晰了不同特征对图书推荐的贡献程度,从高到低依次为作者、关键词、类别、购买者和出版社。

关键词: 图书推荐, 网络表示学习, 图书异构网络